2026 年 10 月 25 日，多家媒体报道称，Replit 已将大部分 AI 支出从 Anthropic 转向 OpenAI。OpenAI 为一份年金额超过 1 亿美元的合同提供了超过 30%的折扣。Rippling、DataCamp 等公司也在增加对 OpenAI 的使用；DataCamp 称，采用 OpenAI 后，公司成本降低了 80%至 90%。
大客户争夺升温的同时，OpenAI 和 Anthropic 近几个月也接连下调模型价格，单款模型的降幅最高达 80%。企业合同折扣与 API 公开定价的调整，正在共同改写大模型市场的价格体系。
Replit 转向 OpenAI：一份年额超 1 亿美元的合同
OpenAI 给出了超过 30% 的折扣，合同年金额超过 1 亿美元——正是这个数字，让 Replit 把大部分 AI 支出从 Anthropic 挪到了 OpenAI。
Anthropic 此前也给 Replit 提供过模型折扣。据报道，Replit 原本计划继续与 Anthropic 合作，但 Anthropic 表示，如果没有更高的消费承诺，他们将不再提供折扣。
不过，首选 OpenAI 和继续支持 Claude 并不冲突。10 月 16 日，CEO 兼联合创始人 Amjad Masad 在 TechCrunch Disrupt 上提到了三项考量：成本、性能、合作关系。他称 OpenAI 是 Replit Agent 的首选合作伙伴。随后又澄清，Replit 仍然保持模型中立，用户可以继续选择 Claude。
Replit 的 Agent 系统会把任务拆给不同模型和子 Agent，日常任务交给便宜的模型，复杂任务交给强的模型。这意味着它的采购结构天然分散：即使大部分支出流向 OpenAI，Claude 仍会留在系统里处理特定任务。Replit 9 月 29 日的工程博客列出了生产环境中的评估对象——Claude Fable 5、Fable 5.1 和 OpenAI Astra。Masad 说“模型中立”是这套架构的必然结果。
折扣换来的是入口。Replit 8 月上线 Free Mode，底层跑的是 OpenAI 的 GPT-5.6 Luna，订阅用户在套餐额度范围内，可以使用它处理日常任务，而无需消耗 credits——没有那轮降价，这个"免费"根本撑不起来。OpenAI 出钱，Replit 出用户，各取所需。
Masad 创办 Replit 的起点，是一次实际的开发挫折。他曾写道，自己在约旦安曼编程时，每次更换电脑都得重新搭建开发环境。后来，他成为 Codecademy 的创始工程师之一，并在 Facebook 负责 JavaScript 基础设施工作；他的团队维护了 Babel 和 Jest 等工具。Replit 源于他希望让开发者能在浏览器中使用编程环境的尝试。这段经历有助于理解，为什么对 Replit 来说，选择模型供应商既是产品决策，也是运营决策：用户通过同一个开发环境接触这些模型，而这个环境正是 Masad 当初希望普及的工具。
对 OpenAI 来说，与 Replit 签下的合同，是其在模型市场赢得的一笔大额客户订单。在这个市场，模型能力并不能单独决定哪家供应商获得工作负载。折扣可以降低平台向订阅用户提供更多 AI 服务的成本，而默认首选模型也能将更多使用量引向提供该模型的供应商。
据报道的交易规模表明，OpenAI 愿意以更低价格换取更大用量，而且这套策略正在奏效。Rippling、Replit 和 DataCamp 等公司正在把更多 AI 支出转向 OpenAI。
Rippling CEO Parker Conrad 表示，公司将 OpenAI 在编程工作负载中的占比从 16% 提高到了 52%，Anthropic 的占比则从 57% 降至 18%。
DataCamp CEO Jonathan Cornelissen 称，OpenAI 的定价让公司成本降低了 80% 至 90%。一家资产超过 70 亿美元的美国银行，以每位用户不到 7 美元的价格，将 ChatGPT 席位扩展到 500 名员工；通常每位用户的价格为 20 美元或更高。
“这是一场比谁先撑不住、先破产的逐底竞赛。这不可持续。”有开发者在 Hacker News 上这样评论。
几个月内，模型价格最多下调 80%
这轮降价不是一次性调整，而是两家公司在几个月里你来我往打出来的。回看时间线，大致是四个回合。
第一回合，6 月底，双方各自亮牌。 OpenAI 发布 GPT-5.6 Sol 和 Luna，Sol 维持 $5 / $30 / $0.50，Luna 专攻低成本场景，定价 $1 / $6 / $0.10。四天后 Anthropic 发布 Claude Sonnet 5，上市价 $2 / $10 / $0.20，并称这是“临时价格”——言下之意，随时可能涨回去。
第二回合，7 月底，OpenAI 先动手。 7 月 30 日，GPT-5.6 Luna 直接降到 $0.20 / $1.20 / $0.02。Anthropic 的回应是把 Sonnet 5 的“临时价格”转为永久定价。
大幅降价后 GPT-5.6 Luna 的模型质量与单项任务成本。图表中越靠左上越好；如果 OpenAI 没有降价 80%，Luna 的表现就会输给 DeepSeek V4 Pro。图表来自 OpenAI。
第三回合，9 月初，战场从标价转到缓存。 Anthropic 发布 Fable 5.1，标价仍是 $10 / $50，但缓存读取成本从 $1 砍到 $0.25，降幅 75%。
尽管缓存成本通常只在定价说明中被一笔带过，但实际最狠：使用 Agent 编写代码时，缓存读取成本通常占总 Token 支出的绝大部分，约为 96% 至 99%。这是因为缓存包含了每一轮对话之前的全部内容。此外，缓存命中率通常在 90% 至 100% 之间，具体取决于大模型推理服务商的基础设施。综合来看，缓存读取成本降低 75%，意味着使用模型的总成本可降低约 70% 至 75%。
OpenAI 一周后发布 GPT-6 Astra，体量和价格都与 Fable 5.1 对标，但缓存成本没动——这一回合，Anthropic 占了便宜。
第四回合，9 月 22 日，同一天互砍。 OpenAI 发布 GPT-6 Sol，定价 $2 / $10 / $0.20，相比折扣期的 GPT-5.6 Sol 输入输出双双减半；同时发布的 GPT-6 Luna 也减半，降到 $0.10 / $0.50 / $0.01。
Anthropic 当天发布 Claude Opus 5.5，并明确表示：“Opus 5.5 的推理服务所需计算资源少于 Opus 5，定价也体现了这一点。”Opus 5.5 的价格为 $4 / $20 / $0.20，整体远低于 Opus 5：输入和输出成本下调 20%，缓存读取成本降幅更大，最终总成本降低了 40% 至 60%。与此同时，它的表现甚至超过了 Fable 5.1。
收尾，9 月 29 日。 GPT-6 Sol 性能不及预期，OpenAI 一周后补发 GPT-6.1 Sol，性能追平 Astra，缓存价格再砍半到 $0.10，与 Anthropic 把缓存设为输入价 5% 的做法对齐。至此，整个系列累计成本降了 75%。
几轮下来，从 6 月 GPT-5.6 系列到现在，热门大模型的使用成本大幅下滑，而且从输出质量看，并没有为了降价牺牲能力。Vercel AI 模型网关的数据显示，每 Token 成本已降至 2026 年 4 月水平的 43.3%；但整体支出仍在增加，因为模型能力提升的速度超过了降价速度。
每个 Token 的成本正在下降，图源：Vercel
为什么能降这么多？
从上面的时间线可以看出，Anthropic 和 OpenAI 之间存在明显的针锋相对。合理推测，两家公司这几次降价是在回应对方的动作。不过，除非有某种外部因素在起作用，否则没有公司会把价格下调 80%。
从去年 11 月份 Agentic LLM 开始高速增长以来，模型越来越像一种大宗商品。在 OpenCode、Zed Agent 这类 Harness 里切换一下，就能让新模型跑你的代码库；通过 OpenRouter 换个 API 地址就能测试。换模型几乎没有成本，厂商就很难建护城河。竞争者也不只 Google 和 Meta——其他厂商能提供足够替代的模型，价格压力自然而来。
“价格战呗。这行没有护城河。再加上中国厂商的竞争，只会一路卷到底。股市会很难看，但对消费者是好事！”
其中，中国开放模型，是这股价格压力里最硬的一股。
OpenRouter 已成为寻找开放权重模型的重要平台
DeepSeek V4 Flash 0420 今年 4 月发布，价格 $0.14 / $0.28 / $0.0028，缓存读取只相当于输入价的 2%——远低于 OpenAI 和 Anthropic 的 5% 至 10%，实际成本比标价还低。OpenAI 7 月把 Luna 降价 80%，很可能就是冲着这类小模型来的。
如今 DeepSeek V4.1 1017 得分从 24 涨到 45，价格几乎没动；今年 5 月，OpenRouter 上有 13 家不同服务商在托管 DeepSeek V4 Flash。如今，托管更新版 1017 的服务商已增至 29 家。
DeepSeek V4.1 Flash 1017 的服务商，按缓存读取成本排序
开发者可以直接指定最低价格，近三十家服务商争着托管同一模型，定价权也就从模型厂商扩展到了整个推理服务市场。Cloudflare 和 DeepInfra 等端点的实际成本与 DeepSeek 相差不到 10%。
Kimi K3.1 Pro 也在逼近闭源高端模型。它在 Artificial Analysis 指数中的得分为 55，Claude Opus 5.5 为 58；Kimi 的输入和输出价格只有 Opus 5.5 的一半，缓存读取价格相同。Kimi K3.1 Pro 与已有 20 多家推理服务商托管的 Kimi K3 规模和架构相同，权重开放后，预计会有更多服务商提供它。模型能力追上来，托管渠道铺开，低价就更容易成为常态。
而最近这轮降价则发生在 OpenAI 和 Anthropic 接近 IPO 之际。
市场传闻称，OpenAI 计划 2027 年上市，Anthropic 可能几个月内就上，两家每年亏损数十亿美元，投资者要的是扭亏计划。按理说这时候不该一直降价——但 IPO 前最值钱的不是利润率，是用户量和收入规模。折扣锁住 Replit 这种大客户，低价把更多开发者拉进自家 API 和订阅体系，先把用量做上去，利润表留到上市故事里再讲。
但降价不等于这门生意会消失。推理服务商能以更低成本提供服务，模型能力又还在快速提升，价格就还有继续下探的空间。OpenRouter 上那 29 家服务商没理由长期托管一个亏钱的模型——不赚钱就停掉，算力转去跑更赚的。所谓“Token 的真实成本”，最后大概率就是边际成本。
对 OpenAI 和 Anthropic 来说，真正的问题不是价格还能降多少，而是在价格降到底之前，能不能靠用户量和收入规模撑到上市那一天。
参考链接：
《OpenAI 开出最高 90%折扣，挖走 Anthropic 大客户》：https://the-decoder.com/openai-offers-discounts-up-to-90-percent-to-poach-anthropics-biggest-customers/
《OpenAI 和 Anthropic 几个月内将大模型价格最多下调 80%，为什么？》：https://minimaxir.com/2026/10/cheap-llms/
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
On October 25, 2026, several media reports stated that Reflit had shifted most of its AI expenditure from Anthropic to Open AI. Open AI offered a discount of over 30 per cent for a contract of over $100 million per year. Companies such as Ripping and DataCamp are also increasing their use of Open AI; DataCamp claims that after the introduction of Open AI, company costs were reduced by 80 to 90 per cent.
While large clients compete for warmth, Open AI and Anthropic have in recent months successively adjusted model prices, with single models falling by up to 80 per cent. The firm contract discount and the adjustment of the API open pricing are jointly rewriting the price system in the LLM market.
Refrit to Open AI: a contract over $100 million per year
Open AI gave a discount of over 30%, with a contract of over $100 million a year -- that's exactly the figure that allowed Repit to move most AI expenditures from Anthropic to Open AI.
Anthropic also provided a model discount to Reprit. It was reported that Reprit had planned to continue working with Anthropic, but Anthropic indicated that they would no longer offer discounts without a higher consumption commitment.
However, the preferred Open AI and continuing support for Claude does not conflict. On October 16, CEO and co-founder Amjad Massad mentioned three considerations on TechCrunch Disrupt: cost, performance, partnership. He claims that Open AI is the preferred partner of Reprit Agent. It was subsequently clarified that Refrit remained neutral and users could continue to choose Claude.
Agent of Refrit will tear tasks to different models and sub-ages, hand over daily tasks to cheap ones, and hand over complex tasks to strong ones. This means that its procurement structure is naturally fragmented: even if most of the expenditure goes to Open AI, Claude will remain in the system to handle specific tasks. The Engineering Blog of Reprit September 29 lists the subjects to be assessed in the production environment - Claude Fable 5, Fable 5.1 and Open AI Astra. Masad says “model neutrality” is the corollary of this architecture.
The discount is the entrance. Refrit goes online in August Free Mode, and on the bottom is Open AI's GPT-5.6 Luna, which subscribers can use within the quota package to handle their daily tasks without consuming credits — without the price cut, this "free" will not last. Open AI pays, Refrit pays the user, takes the needs.
Masad's starting point for the Reprit was a real development setback. He wrote that he had to re-establish the development environment every time he changed his computer while he was in Amman, Jordan. Later, he became one of Codecademy's founding engineers and worked on JavaScript infrastructure on Facebook; his team maintained tools like Babel and Jest. Reprit stems from an attempt to allow developers to use programming environments in browsers. This experience helps to understand why, for Replit, the choice of model suppliers is both product and operational decision-making: users are exposed to the models through the same development environment, which was the tool that Masad had hoped to use.
For Open AI, the contract with Reprit was a large customer order that it won in the model market. In this market, model capacity does not alone determine which supplier receives the work load. Discounts can reduce the cost of the platform providing additional AI services to subscribers, and the default preferred model can channel more usage to suppliers that provide the model.
The scale of the reported transactions suggests that Open AI is willing to offer more at lower prices and that the strategy is working. Companies such as Ripping, Reprit and DataCamp are shifting more AI spending to Open AI.
According to Ripping CEO Parker Conrad, the company increased the share of Open AI in the programming load from 16 per cent to 52 per cent, and that of Anthropic from 57 per cent to 18 per cent.
DataCamp CEO Jonathan Cornelissen claims that Open AI pricing reduced company costs by 80 to 90 per cent. An American bank with more than $7 billion in assets has extended the ChatGPT seat to 500 employees at a price of less than $7 per user; the usual price is $20 or more per user.
“It's a bottom-up race before anyone can survive and go bankrupt. This is not sustainable. There are developers who comment on Hacker News.
In a few months, the price of the model is reduced by 80%.
It's not a one-time adjustment, it's two companies that came to me for months. Look back at the timeline, roughly four rounds.
Round one, end of June, both sides show off. Open AI publish GPT-5.6 Sol and Luna, Sol maintains $5/ $30/ $0.50, Luna specializes in low-cost scenarios, pricing $1/ $6/ $0.10. Four days later, Anthropic publish Claude Sonet 5, with a market value of $2/$10/$0.20, and calling it a “temporary price” — the meaning of which is likely to rise back at any time.
Round two, July, Open AI first. On July 30, GPT-5.6 Luna fell directly to $ 0.20/$1.20/$0.02. Anthropic responded by converting Sonnet 5's “temporary prices” into permanent pricing.
The quality of the GPT-5.6 Luna model and the cost of a single mission were reduced significantly. The higher the top left in the chart, the better; if Open AI does not drop 80%, Luna's performance will be lost to DeepSeek V4Pro. Chart from Open AI.
Round three, early in September, the battlefield moved from bid to cache. Anthropic publish Fable 5.1, still priced at $10/$50, but the cost of cache reading was cut from $1 to $0.25, a 75 per cent reduction.
Although cache costs are usually only carried out in a single price-fixing note, they are the worst: when using Agent to write codes, cache reading costs usually account for the bulk of total Token expenditure, about 96 to 99 per cent. This is because cache contains all the elements before each round of dialogue. In addition, the cache hit rate usually ranges from 90% to 100%, depending on the infrastructure of the LLM reasoning service provider. Taken together, cache has reduced the cost of reading by 75%, which means that the total cost of using models can be reduced by about 70% to 75%.
One week later, publish GPT-6 Astra, both in volume and price, matched Fable 5.1, but cache costs remained intact - this round, Anthropic took advantage.
Round four, September 22, the same day. Open AI publish GPT-6 Sol, pricing $2/$10/$0.20, which halves GPT-5.6 Sol input and output from discount periods; and publish GPT-6 Luna, which halves to $0.10/$0.50/$0.01.
On the same day, publish Claude Opus 5.5 made it clear that “Opus 5.5's reasoning services require less computing resources than Opus 5, as reflected in pricing.” Opus 5.5 prices were $4/$20/$0.20, much lower overall than Opus 5: 20 per cent lower input and output costs, and cache lower reading costs, with lower final total costs of 40 to 60 per cent. At the same time, it even outperformed Fable 5.1.
Closing, September 29th. GPT-6 Sol did not perform as expected, Open AI re-issues GPT-6.1 Sol a week later, and is able to track Astra, cache prices to half to $ 0.10, in line with Anthropic setting cache at 5% input price. So, the cumulative cost of the whole series fell by 75 percent.
Several rounds down, from the GPT-5.6 series in June to the present, the cost of using the hot door LLM has declined significantly and, in terms of output quality, has not sacrificed its ability to reduce prices. Data from the Vercel AI gateway show that the cost per Token has dropped to 43.3 per cent of April 2026 levels; however, overall expenditure is still increasing, as model capacity has increased faster than price reductions.
Cost per Token is declining. Source: Vercel
Why are you dropping so much?
As can be seen from the time line above, there is a clear contrast between Anthropic and Open AI. It is reasonable to assume that these price reductions were in response to each other ' s actions. However, unless some external factor works, no company will reduce the price by 80 per cent.
Since the rapid growth of Agenic LLM began last November, the model has become more and more like a mass commodity. Toggle in OpenCode, Zed Agent, this kind of Harness, so that the new model can run your code library; to be tested by OpenRouter using an API address. The cost of changing models is almost non-existent, making it difficult for manufacturers to build moats. Not only Google and Meta-other manufacturers can provide enough alternative models, but price pressures naturally come.
“Price war. There's no moat in this line. Plus the competition from Chinese producers, it's only going to end. The stock market will be ugly, but good for consumers!"
Among them, the China Open Model is the strongest of these price pressures.
OpenRouter has become an important platform for finding open weight models.
DeepSeek V4 Flash 0420 in April of this year at a price of $0.14/$0.28/$0.0028, cache reads only 2% of the input price - well below 5% to 10% of Open AI and Anthropic and the actual cost is lower than the standard price. Open AI reduced Luna's price by 80% in July, probably for a small model like this.
The DeepSeek V4.1 1017 score has now risen from 24 to 45, and prices have barely moved; in May this year, OpenRouter had 13 different service providers hosting DeepSeek V4 Flash. Today, the number of service providers hosting the updated 1017 has increased to 29.
DeepSeek V4.1 Flesh 1017 service providers in order of cache read cost
The developers can directly specify the lowest price, and nearly 30 service providers compete for hosting the same model, and pricing rights extend from model manufacturers to the entire reasoning service market. The actual cost of peer points such as Cloudflare and DeepInfra is less than 10% different from DeepSeek.
Kimi K3.1 Pro is also approaching a high-end closed source model. Its score in the Artificial Analysis index is 55, Claude Opus 5.5 is 58; Kimi ' s input and output price is only half of Opus 5.5, and cache reads the same price. Kimi K3.1 Pro is the same size and structure as Kimi K3 already hosted by more than 20 reasoned service providers, and more service providers are expected to provide it when the weights are opened. The ability to follow up the model, the channel of hosting is opened, making it easier for low prices to become normal.
The most recent price reduction occurred as Open AI and Anthropic approached IPO.
According to market hearsay, Open AI was listed in 2027, and Anthropic could be up in a few months, with two families losing billions of dollars a year, and investors asking for a loss plan. It would be logical to keep prices down at this point -- but the most valuable thing before IPO is not the profit margin, but the size of the user and the income. The discount locks this big client, bringing more developers into the API and subscription system at a low price, and then puts the amount of it up and leaves the profit statement on the list.
But the price reduction doesn't mean the business will disappear. The reasoning service providers can provide services at a lower cost, and modelling capacity is rising rapidly, leaving room for further downswing. OpenRouter's 29 service providers have no reason to host a model that is short of money for a long time -- to stop making money, to run for more money. The so-called “real cost of Token”, the final probability is marginal cost.
For Open AI and Anthropic, the real question is not how much price can go down, but whether it can last until the end of the day, depending on the size of the user and the income.
References:
Open AI offers a maximum 90% discount and digs out Anthropic Big Customers: https://the-decoder.com/openai-offers-discounts-up-to-cent-to-poach-antropics-biggest-customers/
Open AI and Anthropic reduced the price of LLM by up to 80% in a few months. Why? https://minimaxir.com/2026/10/cheap-llms/
Auto-fetched and translated by AI Atlas. All rights belong to the original author.